Detecting artefacts in video filmed in endoscopy is an important problem for downstream computer-assisted diagnosis. When tackling this problem, one challenge is that the size of an artefact varies in a wide range. The other challenge is that labeling endoscopic images is labor-extensive and is hard to outsource the labeling task to untrained people without the aid of doctors. In this report, we demonstrate how the performance of a Faster R-CNN model can be improved by scaling an image to the right scale before training and testing. The training method overcomes the issue that a convolution neural network trained on one scale barely works when detecting the same category of objects on a different scale. The method is totally independent of the model and can be easily adapted with other models. Besides, it saves time and memory by focusing on the patches that include objects when training the model. The source code? for this report will be made public upon the publishing of my solution. © 2019 CEUR-WS. All rights reserved.